摘要
BP网络是一种非线性映射网络,三层BP网络可以很好地逼近任一连续函数.BP网络用于图像压缩是一个很好的创意.但因其训练时间过长,而没有得到应有的重视.人的视觉特性对图像的低频部分及图像的边缘部分比较敏感.经小波分解后图像的大部分能量集中在低频和图像的边缘部分,即这部分小波系数比较大,其余部分小波系数很小,接近于零.因此提出了小波域BP网络图像压缩方法.该方法是根据小波变换后小波区域重要性的不同而采用不同的压缩比.即不重要的小波系数采用大压缩比的BP网络(隐节点少),重要的系数采用小压缩比网络(隐节点多)或不进行BP压缩而直接编码.并给出Mat-lab仿真程序.
BP neural network is a kind of non-linearity network. By the research of science, researchers discover that three layers BP neural network can approach precisely any continuum function. It is a very nice idea that BP neural network can be used in the image compression domain. But it didn't get recognition that it should have. The characteristic of man's vision is sensitive to the low frequency and the edge part of the image. Afte, wavelets resolving, energies of the image concentrate in the low frequency and edge part, namely this part o; wavelet coefficient is more and greatly, the rest wavelet coefficient is very small, near to zero. To provide perfect compression effect, therefore, they putted forward a kind of means of association of both, wavelet transform domain BP neural network. They putted forward a method of wavelet domain BP network image compression. This method is based on differ compression ratio in differ domain by difference of essentiality, namely, no significant domain use big compression ratio (hidden nodes are small). Significant domain use small compression ratio (hidden nodes are big) or did not use. Moreover, they gave out Matlab program to simulate it.
出处
《光电技术应用》
2007年第1期65-68,共4页
Electro-Optic Technology Application
基金
中国科学院长春光学精密机械与物理研究所创新基金资助项目(ZJ99130B)